Presentation Information
[O15-3]Deep-learning-assisted micromagnetic model development for Ga-doped Nd-Fe-B magnet
*Nikita Kulesh1, Anton Bolyachkin1, Xin Tang1, Hajime Nakamura2, Tadakatsu Ohkubo1, Hossein Sepehri-Amin1 (1. National Institute for Materials Science (Japan), 2. Shin-Etsu Chemical Co., Ltd. (Japan))
Keywords:
Nd-Fe-B magnet,Micromagnetic simulation,Deep learning,FIB-SEM tomography,Image processing
Development of high-performance Nd-Fe-B permanent magnets largely depends on complex microstructural optimization, including control of grain size and texture, engineering of grain boundaries, and control of secondary phases [1]. Modern analytical methods, such as electron microscopy and atom probe tomography, provide multi-scale insights that correlates the performance to microstructure. Micromagnetic simulation is the key tool for evaluating effects of observed microstructural features on coercivity, but it is often limited to oversimplified geometries that lack critical real-world features such as realistic grain shape, size, orientation, and intergranular phase distribution [2]. As control over microstructure improves, incorporating this complexity into models becomes essential for producing insightful simulation outputs. This work is a case study combining modern analytical and computational tools—including FIB-SEM tomography, state-of-the-art deep learning image analysis, and micromagnetic simulations—to investigate a Ga-doped, heavy-rare-earth-free Nd-Fe-B magnet [3] with a coercivity of 2.3 T.FIB-SEM tomography data was acquired with a spatial resolution of 40 nm in all directions, using two imaging modes: back-scattered electron and in-lens. After image pre-processing, spatial phase distribution was determined through contrast-based semantic segmentation of the voxelated volume, utilizing an nnUNetv2 model pretrained on a smaller dataset segmented via k-means clustering. Information on individual Nd2Fe14B grains was obtained via instance segmentation using a NISNet3D model trained on a custom dataset, followed by iterative semi-manual corrections. This approach yielded a comprehensive 3D microstructural dataset containing both phase composition and individual grain information (Fig. 1a). In the presentation, statistical analysis of the resulting 3D data will be discussed, along with comparisons to 2D image-based analyses (Fig. 1b,c).
The segmentation results closely represent the experimental tomography data at the given resolution, making them a suitable direct input for a finite-difference micromagnetic model. Since thin grain boundaries cannot be resolved using SEM microscopy, they were introduced during post-processing. For micromagnetic simulations, the large geometry was divided into a series of smaller patches to accommodate computational constraints. We demonstrate how the developed model can be applied to test various hypotheses, including exchange-coupled vs. exchange-decoupled grains, the influence of weakly magnetic secondary phases, and the effect of texture.In this talk, we will present the proposed pipeline for processing FIB-SEM tomography data, the 3D dataset that can serve as training data for deep learning models, and a micromagnetic model based on the experimental microstructure, along with the corresponding simulation results. Finally, we will discuss our perspective on further optimizing realistic model development to use a single 2D image and on adoption of large-scale simulation strategies.
The segmentation results closely represent the experimental tomography data at the given resolution, making them a suitable direct input for a finite-difference micromagnetic model. Since thin grain boundaries cannot be resolved using SEM microscopy, they were introduced during post-processing. For micromagnetic simulations, the large geometry was divided into a series of smaller patches to accommodate computational constraints. We demonstrate how the developed model can be applied to test various hypotheses, including exchange-coupled vs. exchange-decoupled grains, the influence of weakly magnetic secondary phases, and the effect of texture.In this talk, we will present the proposed pipeline for processing FIB-SEM tomography data, the 3D dataset that can serve as training data for deep learning models, and a micromagnetic model based on the experimental microstructure, along with the corresponding simulation results. Finally, we will discuss our perspective on further optimizing realistic model development to use a single 2D image and on adoption of large-scale simulation strategies.
